ai-bookkeeper reads receipts so they can be matched to bank lines. On 626 real receipts (SROIE 2019) it got 601 of 625 totals right. One cause of the misses: a line like TOTAL PAID RM50.00 printed under GRAND TOTAL RM21.85 and CHANGE RM28.15. The 50.00 is the cash handed over, and the bank would never see it.
The reader now checks: if the paid amount less the change equals an earlier total line (within 5 cents), that line is the total. Card receipts that print only TOTAL PAID are untouched. Result on the same set: 604 of 625 (96.6%), merchant and date unchanged. This was measured on the dataset's word boxes, not on OCR output; I had no tesseract in this session. Estimates only, not accounting advice.